reduce contention on startup index generation (#31006)

runtime/src/in_mem_accounts_index.rs

@@ -103,7 +103,7 @@ pub struct InMemAccountsIndex<T: IndexValue, U: DiskIndexValue + From<T> + Into<
     flushing_active: AtomicBool,
 
     /// info to streamline initial index generation
-    startup_info: Mutex<StartupInfo<T>>,
+    startup_info: StartupInfo<T>,
 
     /// possible evictions for next few slots coming up
     possible_evictions: RwLock<PossibleEvictions<T>>,
@@ -131,9 +131,7 @@ pub enum InsertNewEntryResults {
 }
 
 #[derive(Default, Debug)]
-struct StartupInfo<T: IndexValue> {
-    /// entries to add next time we are flushing to disk
-    insert: Vec<(Slot, Pubkey, T)>,
+struct StartupInfoDuplicates<T: IndexValue> {
     /// entries that were found to have duplicate index entries.
     /// When all entries have been inserted, these can be resolved and held in memory.
     duplicates: Vec<(Slot, Pubkey, T)>,
@@ -141,6 +139,14 @@ struct StartupInfo<T: IndexValue> {
     duplicates_put_on_disk: HashSet<(Slot, Pubkey)>,
 }
 
+#[derive(Default, Debug)]
+struct StartupInfo<T: IndexValue> {
+    /// entries to add next time we are flushing to disk
+    insert: Mutex<Vec<(Slot, Pubkey, T)>>,
+    /// pubkeys with more than 1 entry
+    duplicates: Mutex<StartupInfoDuplicates<T>>,
+}
+
 #[derive(Default, Debug)]
 /// result from scanning in-mem index during flush
 struct FlushScanResult<T> {
@@ -168,7 +174,7 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
             flushing_active: AtomicBool::default(),
             // initialize this to max, to make it clear we have not flushed at age 0, the starting age
             last_age_flushed: AtomicU8::new(Age::MAX),
-            startup_info: Mutex::default(),
+            startup_info: StartupInfo::default(),
             possible_evictions: RwLock::new(PossibleEvictions::new(1)),
             // Spread out the scanning across all ages within the window.
             // This causes us to scan 1/N of the bins each 'Age'
@@ -663,7 +669,7 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
         assert!(self.storage.get_startup());
         assert!(self.bucket.is_some());
 
-        let insert = &mut self.startup_info.lock().unwrap().insert;
+        let mut insert = self.startup_info.insert.lock().unwrap();
         items
             .into_iter()
             .for_each(|(k, v)| insert.push((slot, k, v)));
@@ -1033,7 +1039,7 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
     }
 
     fn write_startup_info_to_disk(&self) {
-        let insert = std::mem::take(&mut self.startup_info.lock().unwrap().insert);
+        let insert = std::mem::take(&mut *self.startup_info.insert.lock().unwrap());
         if insert.is_empty() {
             // nothing to insert for this bin
             return;
@@ -1050,7 +1056,7 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
         drop(map_internal);
 
         // this fn should only be called from a single thread, so holding the lock is fine
-        let mut startup_info = self.startup_info.lock().unwrap();
+        let mut duplicates = self.startup_info.duplicates.lock().unwrap();
 
         // merge all items into the disk index now
         let disk = self.bucket.as_ref().unwrap();
@@ -1062,12 +1068,12 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
                 match current {
                     Some((current_slot_list, ref_count)) => {
                         // already on disk, so remember the new (slot, info) for later
-                        startup_info.duplicates.push((slot, k, entry.1));
+                        duplicates.duplicates.push((slot, k, entry.1));
                         if let Some((slot, _)) = current_slot_list.first() {
                             // accurately account for there being a duplicate for the first entry that was previously added to the disk index.
                             // That entry could not have known yet that it was a duplicate.
                             // It is important to capture each slot with a duplicate because of slot limits applied to clean.
-                            startup_info.duplicates_put_on_disk.insert((*slot, k));
+                            duplicates.duplicates_put_on_disk.insert((*slot, k));
                         }
                         Some((current_slot_list.to_vec(), ref_count))
                     }
@@ -1087,13 +1093,16 @@ impl<T: IndexValue, U: DiskIndexValue + From<T> + Into<T>> InMemAccountsIndex<T,
     /// These were collected for this bin when we did batch inserts in the bg flush threads.
     /// Insert these into the in-mem index, then return the duplicate (Slot, Pubkey)
     pub(crate) fn populate_and_retrieve_duplicate_keys_from_startup(&self) -> Vec<(Slot, Pubkey)> {
-        let mut write = self.startup_info.lock().unwrap();
+        let inserts = self.startup_info.insert.lock().unwrap();
         // in order to return accurate and complete duplicates, we must have nothing left remaining to insert
-        assert!(write.insert.is_empty());
+        assert!(inserts.is_empty());
+        drop(inserts);
+
+        let mut duplicate_items = self.startup_info.duplicates.lock().unwrap();
+        let duplicates = std::mem::take(&mut duplicate_items.duplicates);
+        let duplicates_put_on_disk = std::mem::take(&mut duplicate_items.duplicates_put_on_disk);
+        drop(duplicate_items);
 
-        let duplicates = std::mem::take(&mut write.duplicates);
-        let duplicates_put_on_disk = std::mem::take(&mut write.duplicates_put_on_disk);
-        drop(write);
         duplicates_put_on_disk
             .into_iter()
             .chain(duplicates.into_iter().map(|(slot, key, info)| {